基于高阶微观特征交互的晶圆异常检测研究  

Research on Anomaly Detection Based on High-order Microscopic FeatureInteraction

作　　者：于健博 杨晓峰 YU Jianbo;YANG Xiaofeng(College of Microelectronics,Fudan University,Shanghai 200433,China;Engineering and Applied Technology Research Institute,Fudan University,Shanghai 200433,China)

机构地区：[1]复旦大学微电子学院,上海200433 [2]复旦大学工程与应用技术研究院,上海200433

出　　处：《控制工程》2023年第8期1519-1527,共9页Control Engineering of China

基　　金：国家自然科学基金青年科学基金资助项目(62203118)。

摘　　要：为提高晶圆异常检测准确率,并针对基于分块策略进行异常检测过程中,特征信息无法交互的问题,提出一种基于高阶微观特征交互的晶圆异常检测模型——堆叠交互自编码器(stacked interactive autoencoder, SinAE)。首先,采用层次聚类对采集到的高维数据进行无监督聚类,并根据戴维森堡丁指数(Davies-Boulding index,DBI)的肘拐点确立最优聚类状态;随后,利用Sin AE模型对聚类后的各个数据块分别提取高阶特征,并开发出特征交互模块,对各块高阶微观特征进行维度统一、信息交互以及特征融合操作;然后,对融合交互后的特征进行解码重构以及联立训练;最后,根据实时数据在Sin AE上的损失值,确定数据的异常状态,并采用数值仿真过程数据以及真实的半导体晶圆数据进行算法验证,实验结果表明,所提出的算法具有更高的异常检测性能。To improve the accuracy of wafer anomaly detection and overcome the problem that feature information cannot be interacted with in traditional multiblock-based detection methods,a wafer anomaly detection algorithm based on the interaction of higher-order microscopic features,namely,stacked interactive autoencoder(SinAE)is proposed.Firstly,the collected high-dimensional data is divided by a hierarchical clustering algorithm for unsupervised segmentation,and the optimal clustering status is determined based on the elbow infection point of the Davies-Boulding index(DBI).Then,high-order microscopic features of clustered data blocks are extracted by using the SinAE model,and a feature interaction module is developed to perform dimensional unification,information interaction,and feature fusion on the extracted features.Then,the features after fusion and interaction are decoded,reconstructed and trained simultaneously.Finally,the anomaly status is detected based on the loss values of the real-time data on the SinAE model;the numerical simulation data and real semiconductor wafer data are used to validate the algorithm,and the experimental results show that the proposed algorithm has higher anomaly detection performance.

关 键 词：高阶特征 微观信息 特征交互 异常检测 

分 类 号：TP18[自动化与计算机技术—控制理论与控制工程]